Sarvam AI: Full Capabilities Guide — Models, API, Speech, Vision & How to Run (2026)
Sarvam AI builds India-first LLMs, speech, translation, and document intelligence. Complete guide to Sarvam-30B, Sarvam-105B, Saaras v3, Bulbul v3, Sarvam Vision, API pricing in ₹, and OpenAI-compatible integration.
Sarvam AI is building India's sovereign AI stack — not a single model, but a full product layer spanning chat LLMs, speech recognition, text-to-speech, translation, and document intelligence, all optimized for the way Indian languages are actually used: native script, romanized WhatsApp Hindi, code-mixed Hinglish, and 22 scheduled languages.
In March 2026, Sarvam open-sourced Sarvam-30B and Sarvam-105B — MoE reasoning models trained from scratch on IndiaAI Mission compute. Both are already in production: Sarvam 30B powers Samvaad (conversational agent platform), Sarvam 105B powers Indus (AI assistant for complex reasoning and agentic workflows).
This guide maps every model, API, pricing tier, and integration path — so you can pick the right Sarvam capability for your use case without reading six separate doc pages.
Sarvam AI was founded in 2023 in Bengaluru by Vivek Raghavan and Pratyush Kumar. It was selected under India's IndiaAI Mission to build the country's first homegrown LLM stack — trained entirely on Indian compute with datasets emphasizing Indian languages, code-mixed text, and culturally grounded content.
The strategic bet: unified multimodal models from Western labs treat Indian languages as secondary. Sarvam builds specialized foundations for multilingual India — the same thesis Ideogram applies to design typography, applied here to speech, script diversity, and romanized colloquial usage.
For sovereign AI policy context, see our India Sovereign AI Status 2026 post. This guide focuses on product capabilities and developer integration.
Chat LLMs: Sarvam-30B and Sarvam-105B
Both models are reasoning models trained from scratch — not fine-tunes of Mistral, Qwen, or Llama. Architecture: Mixture-of-Experts Transformer with 128 sparse experts, sigmoid-based routing, and in-house RL (async GRPO with CISPO-inspired policy optimization).
Sarvam 105B leads on agentic benchmarks — BrowseComp and Tau2 — reflecting training on tool interaction, web search, and multi-step environments. On Indian-language pairwise evals, it wins ~90% of comparisons across fluency, script correctness, usefulness, and verbosity.
Sarvam-30B (efficient)
Spec
Value
Total parameters
30B MoE
Active params
2.4B per token
Attention
Grouped Query Attention (GQA)
Context window
64K tokens
Pre-training
16T tokens
License
Apache 2.0
Powers
Samvaad conversational platform
Inference
H100, L40S, Apple Silicon (MXFP4)
Benchmark
Sarvam-30B
Math500
97.0
HumanEval
92.1
LiveCodeBench v6
70.0
AIME 25 (w/ tools)
80.0 (96.7)
BrowseComp
35.5
Indian language win rate
~89% pairwise
Sarvam 30B is optimized for real-time deployment — Sarvam reports 3–6× throughput vs Qwen3 baseline on H100, and runs locally on MacBook Pro M3 via MXFP4.
Choosing between them
Need
Model
Voice-agent pipeline, low latency
Sarvam-30B
Multi-step reasoning, tool use, long docs
Sarvam-105B
Local/edge inference on laptop
Sarvam-30B (MXFP4)
Maximum Indian-language quality
Sarvam-105B
Cost-sensitive high-volume chat
Sarvam-30B (₹2.5/1M input vs ₹4)
API integration (OpenAI-compatible)
python
from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.chat.completions(
model="sarvam-105b",
messages=[
{"role": "user", "content": "Explain GST impact on Indian MSMEs in Hindi."}
],
temperature=0.5,
max_tokens=2000,
)
print(response.choices[0].message.content)
Streaming is supported. Reasoning mode is on by default (reasoning_effort: low) — reasoning tokens count toward max_tokens. Increase max_tokens or set reasoning_effort=None to disable.
Bulbul v3 converts text to natural-sounding speech across Indian languages.
Spec
Value
Model ID
bulbul:v3
Languages
11 (10 Indic + English)
Speakers
30+ voices (Shubh, Priya, Aditya, Ritu, Anand, …)
Max chars
2,500 per REST request
Sample rates
8–48 kHz (48 kHz REST/WebSocket only)
Pace control
0.5×–2.0×
Example
python
from sarvamai import SarvamAI
from sarvamai.play import play
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.text_to_speech.convert(
text="आपका ऑर्डर confirm हो गया है।",
target_language_code="hi-IN",
model="bulbul:v3",
speaker="priya",
speech_sample_rate=24000,
)
play(response)
Critical limitation: Romanized Indic input degrades quality significantly. Always use native script for Indic words — e.g. "आपका order confirm हो गया है" not "Aapka order confirm ho gaya hai".
response = client.text.translate(
input="भारत एक महान देश है।",
source_language_code="hi-IN",
target_language_code="gu-IN",
model="sarvam-translate:v1",
)
Open weights available on Hugging Face under Apache 2.0.
Mayura (colloquial + code-mixed)
python
response = client.text.translate(
input="Your EMI of Rs. 3000 is pending",
source_language_code="en-IN",
target_language_code="hi-IN",
mode="modern-colloquial",
output_script="fully-native",
numerals_format="native",
)
# → "आपका रु. 3000 का ई.एम.ऐ. पेंडिंग है।"
Also available:/transliterate (script conversion without translation) and /detect-language (language ID across all major Indian languages).
Sarvam Vision is a 3B parameter vision-language model built for Indian-language OCR and document parsing — where global VLMs treat Indic scripts as secondary.
Spec
Value
Model ID
sarvam-vision
Parameters
3B (state-space VLM)
Languages
23 (22 Indic + English)
Input
PDF, PNG, JPG, ZIP
Output
HTML, Markdown, JSON (structured page data)
Max pages
10 per job
Max file size
200 MB
Capabilities
Text extraction with layout and reading order preserved
Rate limits: Starter 60 req/min · Pro 200 · Business 1,000 · Enterprise custom.
Free tier: ₹100 credits on signup to explore all APIs.
Products built on Sarvam models
Product
Model
Description
Indus
Sarvam-105B
AI assistant for complex reasoning and agentic workflows
Samvaad
Sarvam-30B
Conversational agent platform for real-time multilingual chat
Both are live in production — the open-source release is not a research preview; these models serve real users today.
Sarvam Startup Program (March 2026): Selected early-stage companies receive 6–12 months of API credits, priority engineering support, and production infrastructure access.
Building a voice agent pipeline
The most common production pattern stacks three Sarvam APIs:
Why Sarvam-30B for the LLM layer: 2.4B active parameters = low latency; 64K context handles conversation history; trained on code-mixed Indian language input natively.
For agentic voice (tool calling, web search): swap in Sarvam-105B — 49.5 BrowseComp and 68.3 Tau2 scores reflect strong tool-use training.
For document-heavy workflows (scan a form, extract fields, respond in voice): add Sarvam Vision upstream of the LLM.
Honest benchmark framing
Sarvam's strength is structural, not universal frontier dominance:
Where Sarvam leads:
Indian-language pairwise evals (~90% win rate for 105B)
Agentic benchmarks in its class (Tau2, BrowseComp)
Math/reasoning at model scale (Math500 98.6, AIME 96.7 w/ tools)
Tokenizer efficiency for Indic scripts (lower cost per Indic token)
Speech/translation/OCR for 22+ languages
Where Sarvam trails:
English-centric global frontier benchmarks (Artificial Analysis Intelligence Index ~18 for 105B)
TerminalBench Hard (~1.5% for 105B vs GLM-4.5-Air ~20%)
SWE-Bench Verified (45% — competitive but below top coding models)
The honest use case: Indian-language applications, voice agents, document digitization, and sovereign deployment — not replacing Claude Fable 5 for English-only frontier coding.
MCP and agent integration
Sarvam publishes an MCP server at https://docs.sarvam.ai/_mcp/server for Claude Code, Cursor, and other MCP hosts — plus a Meta Prompt in their docs to guide any chat model on using Sarvam APIs effectively.
For wiring into agent harnesses, Sarvam-105B's tool-use training (BrowseComp 49.5, Tau2 68.3) makes it a strong backend for Indian-language agent loops. See our Agent Harness guide for loop architecture.
Test in Playground at docs.sarvam.ai before production
For self-hosted LLM: download weights from Hugging Face, run with vLLM/SGLang
For voice agents: Saaras → Sarvam-30B → Bulbul pipeline
For documents: Sarvam Vision batch API (split PDFs >10 pages)
Summary
Sarvam AI is the most complete India-first AI product stack available in 2026 — not just LLMs, but speech, translation, TTS, and document intelligence trained on Indian compute with open weights on the flagship models.
Three things to remember:
Two LLMs, two jobs: Sarvam-30B for speed and voice pipelines; Sarvam-105B for reasoning, agents, and maximum quality.
Translation has two modes: Sarvam-Translate for formal 22-language coverage; Mayura for colloquial and code-mixed Hinglish.
The moat is Indic depth — ~90% win rate on Indian-language benchmarks, native-script OCR, and code-mixed speech — not English frontier parity.
Model specs, API pricing, and benchmark numbers reflect Sarvam's public documentation as of June 20, 2026. Verify current pricing and model availability at docs.sarvam.ai before production deployment.
